Merge pull request #747 from NeptuneHub/devel

AI tools integration fix

Author: NeptuneHub

tasks/ai/prompts.py

@@ -284,13 +284,39 @@ def build_ollama_tool_calling_prompt(
 Fill only fields the user asked for. Return ONLY the JSON object."""
 
 
+def _inject_unique_items(schema: Dict) -> Dict:
+    """Add ``uniqueItems: True`` to every array-typed property recursively.
+
+    Small Ollama models can loop the same value forever in structured-output
+    mode; ``uniqueItems`` prevents that.  Injected here rather than in the
+    shared ``tools.py`` schema so that only the Ollama structured-output path
+    is affected -- Gemini, Mistral, and native OpenAI tool-calling use the
+    schema as-is.
+    """
+    if not isinstance(schema, dict):
+        return schema
+    # Snapshot keys so we can safely mutate the dict while walking it.
+    for key, value in list(schema.items()):
+        if key == 'type' and value == 'array':
+            schema.setdefault('uniqueItems', True)
+        elif isinstance(value, dict):
+            _inject_unique_items(value)
+        elif isinstance(value, list):
+            for item in value:
+                if isinstance(item, dict):
+                    _inject_unique_items(item)
+    return schema
+
+
 def build_tool_calls_schema(tools: List[Dict]) -> Dict:
     branches: List[Dict] = []
     for t in tools:
         name = t.get('name')
         if not name:
             continue
-        arg_schema = copy.deepcopy(t.get('inputSchema') or {"type": "object"})
+        arg_schema = _inject_unique_items(
+            copy.deepcopy(t.get('inputSchema') or {"type": "object"})
+        )
         arg_schema['additionalProperties'] = False
         branches.append(
             {

tasks/ai/tools.py

@@ -16,7 +16,7 @@
 Main Features:
 * get_mcp_tools builds the schema dynamically, exposing text_match modes only when CLAP/LYRICS are enabled; tool descriptions carry the routing rules (when to use each tool and when to use a sibling instead) so they work as the primary routing signal for small models, with genre/voice/mood enums from the canonical vocab.
 * execute_mcp_tool converts normalized energy 0..1 to raw score units before search_database, expands female/male voice spelling variants deterministically, passes exclude_artists/exclude_genres through as hard SQL cuts, and rejects year-only text_match queries (routing them to search_database); all failures return a generic error, never a traceback.
-* Array args carry maxItems/uniqueItems caps so small-model structured output cannot loop a value forever; exclusion fields document that excluded names never go in seeds or positive filters.
+* Array args carry maxItems caps so small-model structured output cannot loop a value forever; the Ollama structured-output path in prompts.py adds uniqueItems dynamically. Exclusion fields document that excluded names never go in seeds or positive filters.
 """
 
 import logging
@@ -493,14 +493,12 @@ def get_mcp_tools() -> List[Dict]:
                         "type": "array",
                         "items": {"type": "string", "enum": list(GENRE_VOCAB)},
                         "maxItems": 5,
-                        "uniqueItems": True,
                         "description": "Music genres the user WANTS, e.g. ['jazz'] or ['rock', 'blues'].",
                     },
                     "voices": {
                         "type": "array",
                         "items": {"type": "string", "enum": list(VOICE_ENUM)},
                         "maxItems": 2,
-                        "uniqueItems": True,
                         "description": (
                             "Vocal type: 'female vocalists' for any female-voice request, "
                             "'male vocalists' for any male-voice request."
@@ -510,7 +508,6 @@ def get_mcp_tools() -> List[Dict]:
                         "type": "array",
                         "items": {"type": "string", "enum": list(config.OTHER_FEATURE_LABELS)},
                         "maxItems": 3,
-                        "uniqueItems": True,
                         "description": "How it feels, e.g. ['sad'] or ['danceable', 'party'].",
                     },
                     "tempo_min": {"type": "number", "description": "Min BPM 40-200, e.g. 120"},
@@ -552,7 +549,6 @@ def get_mcp_tools() -> List[Dict]:
                         "type": "array",
                         "items": {"type": "string"},
                         "maxItems": 10,
-                        "uniqueItems": True,
                         "description": (
                             "Artists the user does NOT want ('no 50 Cent' -> ['50 Cent']). "
                             "Hard-removed from the results; never put these in artist or seeds."
@@ -562,7 +558,6 @@ def get_mcp_tools() -> List[Dict]:
                         "type": "array",
                         "items": {"type": "string", "enum": list(GENRE_VOCAB)},
                         "maxItems": 5,
-                        "uniqueItems": True,
                         "description": (
                             "Genres the user does NOT want ('no rap' -> ['Hip-Hop']). "
                             "Hard-removed from the results; never put these in genres."

test/unit/test_mcp_server.py

@@ -607,6 +607,21 @@ def test_unknown_tool_returns_error(self):
 
 
 class TestToolSurface:
+    def test_shared_schemas_omit_ollama_only_unique_items(self):
+        ai_mod = _import_ai_mcp_client()
+
+        def contains_unique_items(value):
+            if isinstance(value, dict):
+                return 'uniqueItems' in value or any(
+                    contains_unique_items(child) for child in value.values()
+                )
+            if isinstance(value, list):
+                return any(contains_unique_items(child) for child in value)
+            return False
+
+        for tool in ai_mod.get_mcp_tools():
+            assert not contains_unique_items(tool['inputSchema'])
+
     def test_no_llm_facing_get_songs_or_dead_params(self):
         ai_mod = _import_ai_mcp_client()
         for tool in ai_mod.get_mcp_tools():

test/unit/test_tool_plan.py

@@ -593,6 +593,39 @@ def test_source_schema_not_mutated(self):
         pr.build_tool_calls_schema(tools)
         assert 'additionalProperties' not in tools[0]['inputSchema']
 
+    def test_unique_items_are_added_only_to_derived_argument_arrays(self):
+        pr = _prompts()
+        tools = _tools_fixture()
+
+        schema = pr.build_tool_calls_schema(tools)
+        branches = schema['properties']['tool_calls']['items']['oneOf']
+        argument_schemas = [branch['properties']['arguments'] for branch in branches]
+
+        def array_schemas(value):
+            if isinstance(value, dict):
+                if value.get('type') == 'array':
+                    yield value
+                for child in value.values():
+                    yield from array_schemas(child)
+            elif isinstance(value, list):
+                for child in value:
+                    yield from array_schemas(child)
+
+        derived_arrays = [
+            array_schema
+            for argument_schema in argument_schemas
+            for array_schema in array_schemas(argument_schema)
+        ]
+        source_arrays = [
+            array_schema
+            for tool in tools
+            for array_schema in array_schemas(tool['inputSchema'])
+        ]
+
+        assert derived_arrays
+        assert all(array_schema['uniqueItems'] is True for array_schema in derived_arrays)
+        assert all('uniqueItems' not in array_schema for array_schema in source_arrays)
+
 
 class TestPromptRendering:
     def test_system_prompt_derives_tools_and_rules(self):